Human physiological time series data, including electrocardiograms (ECGs) and wearable sensors data, provide essential information for healthcare applications. The inherent complexity and variability of these signals introduce both challenges and opportunities for deep learning research. This thesis systematically addresses several critical aspects of physiological time series analysis:

(1) We begin by empirically studying the behavior of the "Mixup" family of data augmentation methods, traditionally employed in computer vision, in the context of physiological time series. Through extensive experiments, we verify the unique properties of Mixup in this domain, extending known techniques to new applications.

(2) We introduce a novel two-stage curriculum learning approach for seizure forecasting, a critical task in epilepsy treatment. By employing an autoencoder-based scoring stage followed by a subsequent fine-tuning stage involving a pre-trained encoder and a ranked sample pool, we address the challenge of class imbalance in forecasting anomaly events. This scheduled sampling scheme demonstrates superior performance and faster model convergence in seizure prediction.

(3) We present a Transformer-based model specifically designed for ECG data analysis. Drawing inspiration from the inherent physiological components of ECG signals, the model adopts a representation learning approach, segmenting ECG signals into distinct fragments or "tokens." These tokens, characterized by physiological significance such as P/QRS/T components in the waveforms, are processed by the Transformer. Experiments demonstrated competitive performance in the arrhythmia classification task. This approach ensures that the model's learning process aligns closely with clinical insights, enhancing its interpretability.

We evaluate the proposed methods on various datasets, including standard 12-lead ECG signals for project 1 and 3, multiple biomedical time series datasets for project 1, and wristband-recorded physiological signals for project 2. The evaluations show promising results and significant improvements compared to baseline models. %Together, Collectively, the projects form a cohesive narrative that reflects a progression from empirical exploration to the design of deep learning models and schemes, all centered around the unique challenges and opportunities of physiological time series modeling.